Anthropic: small models may soon self-replicate

What happened: Anthropic’s Frontier Red Team published experiments suggesting that small AI models may be months away from self-replicating on commercial internet-of-things devices like smart ovens and off-brand Android TV boxes. In its evaluations, a 9B model got inference running on 8 of the 9 hardware types tested, and the team ran a 26B model (Gemma4-26B-A4B) sharded across 15 used Android TV boxes bought for about $55 each. It also showed a model spreading node to node across a firewalled fleet of Android emulators.

Context: Anthropic says models of 30B parameters and under can already complete many of the individual replication steps but make mistakes when composing them end to end, while GLM-5.2 managed full end-to-end replication in simplified evaluations, spreading to 8 of 8 nodes in four of five trials. It calls these weight-based replicators, since the model weights themselves become the malware payload, and notes they need no root access and leave no central switch to flip, though they would spread slowly, with a doubling time measured in days. If small models keep improving at their historical rate, Anthropic projects one will reach that level by the end of 2026.

Why it matters: This is a rare, measured public look at the “AI worm” idea, with a concrete capability estimate instead of a hypothetical. Anthropic also sketches the defenses, from traffic anomaly detection and ISP quarantines to secure-by-default devices, and argues the same coding ability that makes agents useful is what makes replication possible.

(Original source: Anthropic)